Deep learning the slow modes for rare events sampling.

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Machine Learning

Background:

• Enhanced sampling methods in atomistic simulations require accurate collective variables (CVs) to study long-time phenomena.
• Identifying optimal CVs often relies on prior knowledge of system dynamics, which is typically unavailable.

Purpose of the Study:

• To develop a robust algorithm for automatically identifying efficient collective variables from biased simulations.
• To accelerate the sampling of rare events in molecular systems using machine learning and enhanced sampling techniques.

Main Methods:

• Utilizes machine learning, specifically a neural network ansatz, to extract transfer operator eigenfunctions.
• Integrates the on-the-fly probability-enhanced sampling method for efficient CV identification and acceleration.
• Applies the algorithm to diverse systems including small molecule transitions, protein folding, and materials crystallization.

Main Results:

• Demonstrates a powerful and robust algorithm capable of extracting relevant CVs from biased simulations.
• Successfully accelerates the sampling of rare events across various molecular and materials systems.
• Validates the generality and effectiveness of the machine learning-driven approach.

Conclusions:

• The developed algorithm automates the identification of collective variables, overcoming limitations of traditional methods.
• This approach significantly enhances the efficiency and scope of atomistic simulations for studying complex phenomena.
• Machine learning combined with enhanced sampling offers a powerful strategy for molecular discovery.